Why do I hit my Claude Code limit so fast?

Because Claude rerereads the entire conversation from message 1 on every single reply. A 30-message thread costs roughly 31 times more per message than a fresh chat. Add MCP servers (each one can load 18,000+ tokens per message), a bloated CLAUDE.md, and pasted files you don't need, and you can burn through your session limit before finishing a single task. The problem is context bloat, not plan size.

What is the fastest way to reduce Claude Code token usage?

Use /clear between unrelated tasks, disconnect MCP servers you're not actively using, and keep your CLAUDE.md under 200 lines. These three changes alone can cut your per-session spend by 50% or more, based on what we've seen at Kreante.

Does running /compact actually help?

Yes, but timing matters. Auto-compact triggers at 95%, by which point your context quality is already degraded. Run it manually at around 60% and give it specific instructions on what to preserve. After 3 to 4 compacts in a row, do a full session summary and use /clear instead.

How do MCP servers affect Claude Code token limits?

Each connected MCP server loads all its tool definitions into your context on every message. One server alone can add 18,000 tokens per message before you've typed anything. If you have 8 to 10 servers connected but only need 2 for a given task, you're burning through your allocation on tools you aren't using.

Should teams use Claude Code for sales and project management tasks?

Yes, and it works well when sessions are structured correctly. The biggest mistake teams make is running everything in one long thread: research, drafts, follow-ups, all piled together. Breaking tasks into focused sessions with /clear between them keeps costs manageable and output quality higher.

18 Claude Code Token Hacks: How We Cut Daily AI Spend Across Sales, Marketing, and PM Teams

The real reason you’re running out of tokens

Claude Code rerereads your entire conversation from the beginning on every single message.

Message 1 costs roughly 500 tokens. Message 30 costs around 15,500, because Claude is processing all 29 previous exchanges before getting to yours. One developer tracked a 100-plus message session and found 98.5% of all tokens went to rereading old history — not doing actual work.

On top of that, every message reloads your CLAUDE.md file, every connected MCP server, your system prompts, and your skills. One MCP server alone can add 18,000 tokens per message before you’ve typed a word.

There’s also a model behavior worth knowing: attention degrades in the middle of long sessions. Models focus most on the beginning and the end. Everything in the middle gets processed with less weight. A bloated context doesn’t just cost more, it produces worse output.

The framework below breaks 18 hacks into three tiers. At Kreante, we applied these across three departments that use Claude daily: sales, marketing, and project management. The examples are from what actually changed in our workflow.

Tier 1: The basics (9 hacks everyone should use)

1. Use /clear between unrelated tasks

Every message in a long session is exponentially more expensive than the same message in a fresh one. Switching from a lead brief to an article draft without clearing context means Claude is carrying sales research into a content task. It wastes tokens and subtly degrades quality.

Our sales team used to run everything in one thread: company research, qualifying notes, email draft, follow-up. Now each prospect gets a fresh session. The briefs got sharper and shorter at the same time.

2. Disconnect MCP servers you’re not using

Run /mcp at the start of each session and check what’s connected. If you have Gmail, Google Calendar, Slack, Figma, Webflow, Fireflies, and n8n all loaded, every single message pays for all of them, even if you’re just writing a doc.

Disconnect what you don’t need for the current task. If you can find a CLI alternative to an MCP server, use it. CLIs are faster and don’t carry token overhead.

3. Batch your prompts into one message

Three messages cost roughly three times what one combined message costs. Instead of “summarize this,” then “extract the issues,” then “suggest a fix” as three separate messages, send all three at once.

If Claude gets something slightly wrong, edit your original message and regenerate. Follow-up corrections stack onto the conversation history permanently. Edits replace the bad exchange entirely.

4. Use plan mode before complex tasks

The biggest single source of token waste is Claude going down the wrong path, writing hundreds of lines of code or copy, and then having to scrap it. Plan mode lets Claude map out the approach and ask clarifying questions before touching anything.

Consider adding this to your CLAUDE.md: “Do not make any changes until you have 95% confidence in what needs to be built. Ask follow-up questions until you reach that level.”

5. Run /context and /cost regularly

These two commands make invisible spending visible. /context shows what’s eating your tokens: conversation history, MCP overhead, loaded files. /cost shows actual usage and estimated spend for the session.

Most people have no idea where their tokens are going. If you don’t know you’re bleeding 18,000 tokens per message on an MCP server you’re not using, you can’t fix it.

6. Set up a status line

Run /status line in your terminal. It adds a persistent bar showing your current model, a visual context usage bar, and your token count. Having this visible while you work changes behavior. You notice when a session is getting heavy before it becomes a problem.

7. Keep your usage dashboard open

Same principle. Knowing your remaining allocation and reset time lets you pace yourself. If you’re near a reset with budget remaining, that’s the time to push hard. If you’re near your limit with hours to go, step away and come back with a full allocation.

8. Paste only what Claude needs

Before dropping a file or document, ask yourself whether Claude needs the whole thing. If the bug is in one function, paste that function. If the context is one paragraph, paste that paragraph.

Our marketing team used to paste full SEO reference guides, competitor analyses, and keyword lists into content sessions. Trimming to the relevant sections cut context weight significantly without changing output quality.

9. Watch Claude work on complex tasks

Don’t fire off a prompt and switch tabs. Watch the first few steps. If Claude starts going down the wrong path, reading irrelevant files, or getting stuck in a loop, stop it early. The cost of catching a bad direction at message 3 versus message 20 is substantial.

Tier 2: Getting surgical (5 hacks)

10. Keep CLAUDE.md under 200 lines

This file reloads on every single message, not just at session start. If your CLAUDE.md is 1,000 lines, you pay for all 1,000 on every exchange, even a one-word reply.

Treat it as an index. Instead of including your full coding conventions, tech stack docs, and process guides inline, write a one-line pointer to where each document lives. Claude knows exactly where to look when it needs something, without loading it preemptively every time.

Our content CLAUDE.md had grown to 800-plus lines of SEO guidelines, style rules, and reference lists. Trimming it to a lean index under 150 lines, with pointers to external files, cut our per-article session overhead noticeably.

11. Be specific with file references

“Here’s my whole repo, find the bug” makes Claude explore freely and read everything it can find. “Check the verifyUser function in auth.js” costs a fraction of that.

Use @filename to point directly at specific files. If you’re asking about a project update, reference the specific section of the project context note, not the full file.

This was a real shift for our project managers. They used to paste entire Slack threads, full project files, and budget summaries into status update sessions. Now they paste the specific section relevant to the question.

12. Compact manually at 60% context

Auto-compact triggers at 95%. By that point your context is already degraded and some of what was in the middle has lost model attention. Run /compact manually at around 60%, with specific instructions on what to preserve.

After 3 to 4 compacts in sequence, quality starts to slip. At that point, get a session summary, use /clear, paste the summary back, and continue in a fresh session.

13. Short breaks cost you

Claude Code uses prompt caching to avoid reprocessing unchanged context. The cache expires after 5 minutes. If you step away and come back after 5 minutes, your next message reprocesses everything from scratch at full cost.

Before you take a break, run /compact or /clear. It’s a small habit with a real impact on sessions that span hours.

14. Watch command output size

When Claude runs shell commands, the full output enters your context window. A git log that returns 200 commits, a build command with verbose output, or a database query with thousands of rows — all of it becomes tokens.

Be intentional about which commands you let Claude run in a given project. In projects where certain commands generate bloated output, you can deny those permissions explicitly.

Tier 3: Advanced (4 hacks)

15. Pick the right model for each task

Sonnet handles most work well. Use Haiku for sub-agents, formatting tasks, research summaries, and anything where you need volume over depth. Reserve Opus for deep architectural planning, and keep it under 20% of total usage.

When you need a large codebase reviewed, consider bringing in a separate tool for that specific step rather than running the review through your main Claude session.

16. Know the real cost of sub-agents

Agent workflows use 7 to 10 times more tokens than a single-agent session. Each sub-agent wakes up with its own full context, reloads everything from scratch, and operates as a separate instance. It adds up fast.

Sub-agents make sense for isolated one-off tasks, especially when you can run them on Haiku. Delegate research summaries, formatting passes, or data extraction to sub-agents on cheaper models rather than processing everything through an expensive main session. Agent teams produce high-quality output but should be used sparingly.

17. Understand peak hours

Anthropic adjusts how fast your session window drains based on platform demand. Peak hours (8 AM to 2 PM Eastern, weekdays) drain faster. Off-peak hours (afternoons, evenings, weekends) run at normal or slightly longer rates.

If you have large refactors, multi-agent workflows, or heavy research sessions, schedule them off-peak. The work is the same; the cost against your session window is lower.

Corollary: if you’re near a session reset and still have budget, that’s the time to push. Let your agents run, tackle the heavy tasks, and get full value from your allocation before it resets.

18. Make CLAUDE.md self-learning

This is the most advanced and the most fragile. Add a section at the bottom of your CLAUDE.md for applied learning: when a workaround gets discovered, when something fails repeatedly, when a platform has a known quirk, add a one-line bullet. Keep each bullet under 15 words. No explanations, just the actionable fact.

Check it regularly to keep it from bloating. But a CLAUDE.md that accumulates real operational learnings from your team’s usage gets progressively more efficient over sessions, not less.

How this played out at Kreante

We run Claude across three departments with different use patterns.

Sales: Lead qualification, prospect brief preparation, and email drafts. The biggest waste was long threads that carried context from one prospect into another. We now use /clear between each prospect. Sessions are shorter, briefs are more focused, and the account managers stopped complaining about weird cross-contamination in the suggestions.

Marketing: Article creation, keyword research, and content iteration. Our CLAUDE.md for content had grown into a catch-all document with full style guides, SEO checklists, and topic matrices loaded inline. Converting it to a pointer index was the single highest-impact change. We also stopped pasting full reference articles and learned to paste only the specific section Claude actually needs.

Project management: Status updates, client communication, and sprint documentation. The change here was surgical pasting. PMs used to load full project context notes, Slack thread exports, and meeting summaries into every session. Now they paste only the section relevant to the specific question. The output quality stayed the same. The session length dropped.

The bottom line

Most people don’t need a bigger Claude plan. They need to stop sending their entire conversation history 30 times when 5 would do the work.

Every hack in this list traces back to the same principle: Claude pays for everything it reads, every time it reads it. The less you make it carry, the more efficient each message becomes.

Start with /context to see what’s actually consuming your tokens. Fix the top item. Then move down the list.